Problem

When a SQL error aborted the transaction inside claim_one, the outer transaction was also rolled
back, preventing the job row from being marked as failed.

Solution

• Wrap the yield p in a nested savepoint transaction (txn.conn.transaction()) so that inner
  aborts don't poison the outer transaction
• After updating the status, re-fetch the event row so p.evt reflects the final DB state

Testing

Added a test that verifies a job is correctly marked failed when the body raises a SQL error (calling
a non-existent function).